cert-machine · audit · an AI model’s proposals, stated by claimbox, decided here

The real roots of xˣ = y: every fraction bracket an AI model proposed, decided exactly

xˣ = y, by fractions

It began with one picture, xˣ = 64. The equation has one real root, and no fraction equals it. What a fraction can do is trap it: one fraction on either side, each side one comparison of two whole numbers, decided exactly. An AI model computed the root for every y to a thousand and proposed the traps; claimbox stated them as claims; this page decides every side again with code of its own and reads what the fractions show.

Decided here: every side of every trap the study proposed, the four whole-number roots from both sides, and the eight claims about xˣ = 64, each by one exact comparison of integers (certs/xx-ledger.json, tools/run-xx-ledger.py). claimbox’s words are shown beside ours. Computed and not decided: the roots’ digits, their continued fractions, the closest fractions and every statistic. Proved in Lean and checked by two kernels: that the root is unique and that a fraction bounds it exactly when one integer exceeds another (certs/xx-reduction-lean.json). Read and not decided: the negative-base convention. Refused: one claim of our own, that the root of 64 is transcendental. Cited and not decided: the Gelfond–Schneider theorem.

tl;dr
  • The finding. 210 of 218 claims CERTIFIED and the other 8, the planted wrong brackets, REFUTED: 105 roots of xˣ = y each trapped between two fractions, and at y = 4, 27, 256 and 3,125 the root pinned to a whole number from both sides. The root of 64 is decided to nine decimals. claimbox’s record gives the same word on every claim. One claim of cert-machine’s own, that the root of 64 is transcendental, is REFUSED: it is true, by a cited theorem, and no comparison of two integers decides it.
  • The mechanism. For a fraction p/q above 1, x > p/q exactly when pp < Nq·qp, where N = y is read from the pinned file; both sides are computed as whole numbers and compared. No logarithm is evaluated. 216 mutants, each fraction moved one unit the wrong way or each whole-number bound made strict, must flip to REFUTED, and all do. That the comparison decides the bracket is proved in Lean and checked by two kernels (§4).
  • Check it. python3 -I tools/run-xx-ledger.py --check (seconds; the Python standard library) re-hashes the copies and re-decides every claim.
roots computed
1,002
one for every y from 2 to 1,000, and three beside 5⁵, by Newton’s method in 160-digit arithmetic
roots trapped
105
each between two fractions, every side one exact comparison of whole numbers
claims certified
210 of 218
the other 8 are the planted wrong brackets, all refuted
roots that are fractions
4
exactly the y of the form kᵏ; every other root is transcendental (a theorem cited)
§1 · the picture

The root of 64, nine decimals decided

x = ln 64 / W(ln 64), where W is Lambert’s function. That form describes the number; it decides nothing. The decimals below are decided by two fractions.

whatthe numberdecided hereclaimbox’s record
the root, computed3.39912154005228331235329865137612775126497803353051——
below itx > 458912/135009CERTIFIEDCERTIFIED
above itx < 400097/117706CERTIFIEDCERTIFIED
so, printed outward3.399121540 < x < 3.399121541CERTIFIED—

xˣ grows for x above 1, so x is above a fraction p/q exactly when (p/q)p/q is below 64, that is, when pp is below 64q·qp. For the fraction below, each side of that comparison is a whole number of 2,598,236 digits. The recorder computes both exactly and compares them; the step from x to the comparison is proved in Lean and checked by two kernels (§4).

458912458912 < 64135009 · 135009458912

The two fractions are neighbouring convergents of the root’s continued fraction, [3; 2, 1, 1, 44, 1, 14, 4, 1, 6, …]: the closest fractions for their size, on opposite sides of x. Its large fifth quotient is why 775/228, with a denominator below a thousand, is already off by only 1.266965e-06.

the claim about xˣ = 64origineach side, digitsdecided hereclaimbox’s record
x > 33991/10000proposed154,026 and 154,026CERTIFIEDCERTIFIED
x < 4249/1250proposed15,417 and 15,417CERTIFIEDCERTIFIED
x > 4249/1250planted wrong15,417 and 15,417REFUTEDREFUTED
x < 33991/10000planted wrong154,026 and 154,026REFUTEDREFUTED
x > 58815/17303proposed280,518 and 280,518CERTIFIEDCERTIFIED
x < 47207/13888proposed220,646 and 220,646CERTIFIEDCERTIFIED
x > 458912/135009proposed2,598,236 and 2,598,236CERTIFIEDCERTIFIED
x < 400097/117706proposed2,241,410 and 2,241,410CERTIFIEDCERTIFIED
x is transcendentalcert-machine’s ownno comparisonREFUSED—

The first two rows are the AI’s four-decimal proposal; the planted rows move it one unit in the last decimal, the wrong way, and are refuted. A refutation here is a comparison of two numbers, so it has no witness point to draw.

The last row is cert-machine’s own claim, and it is refused. This decider answers one question: does the root lie above or below a given fraction, by one comparison of two whole numbers. "Transcendental" says the root is a root of no polynomial with integer coefficients: infinitely many polynomials, not one comparison, and no number of decided brackets reaches it. It is true: the root is not a fraction (p/q in lowest terms with q > 1 would give pp = 64q·qp, which q cannot divide, and no whole k has kk = 64), and an irrational algebraic x would make xˣ transcendental by the Gelfond–Schneider theorem, while 64 is not. That theorem is cited, not decided here (§3, one). What would make it decidable. A kernel-checked proof: the Gelfond–Schneider theorem formalized (Mathlib at the pin of lean/xx-reduction, d13f23b7, holds no Gelfond–Schneider theorem), then the two-line argument above in Lean on top of XXReduction.existsUnique_root. With that, it is a theorem to check, not a comparison to compute. claimbox stated no such claim; it is the one row of this page that is not claimbox’s.

decided
exact arithmetic backs it
computed
float, or a number someone printed; nothing decided it
refused
no mark — the void is drawn
why these numbers stand as they do
  • computed — computed by the proposer’s Newton iteration in 160-digit arithmetic; nothing decided it
  • decided — p, the numerator of the decided fraction
  • decided — N, read from the pinned file
  • computed — a partial quotient of the computed root, kept where the 120- and 160-digit roots agree; computed, not decided
  • refused — refused: not a comparison of the root with one fraction, the one statement this decider reads
§2 · every root

A thousand roots on one curve

Each dot is the root of xˣ = y. A solid dot is trapped between fractions decided here; a hollow dashed one is computed only. The large dots are the roots that are whole numbers.

1 10 100 1,000 1 2 3 4 5 6 x y = xˣ 2² = 4 3³ = 27 4⁴ = 256 5⁵ = 3,125
Solid: the 105 decided roots, each drawn as its bracket: a dot where the two fractions are closer than a pixel, a bar from one to the other where they are not (y = 23, from 38/13 to 3; y = 3,124, from 4 to 5; y = 3,126, from 5 to 6). Dashed: the curve y = xˣ, drawn in floating point, and the 897 roots computed only. The guides fall at x = 2, 3, 4, 5.Records: corpus/claimbox-xx/records/xxy-study.json · sha256 bf49436db85d; certs/xx-ledger.json · sha256 bedd4f1ade74The build refuses this figure if a decided dot’s bracket is not CERTIFIED on both sides here, or claimbox’s word on it differs.

How tight each trap is

For y up to 100, the bar is the number of decimal places on which the two decided fractions agree: those decimals of the root are decided. The dashed line is how many decimals the closest fraction with a denominator of at most 1,000 shares with the computed root. Where the root is a whole number, the bar reaches the top line, labelled exact.

0 2 4 6 8 exact y decimals 2 10 27 50 64 100 0 decimals shared
The bracket’s two fractions are the last two convergents with a denominator of at most one thousand. Where the next partial quotient is large, the closest fraction is far better than its size suggests, and the dashed line rises above the bar.Records: certs/xx-ledger.json · sha256 bedd4f1ade74; corpus/claimbox-xx/records/xxy-study.json · sha256 bf49436db85dThe build refuses this figure if a closest fraction shares fewer decimals with the computed root than its decided bracket pins.

Every continued fraction, at a glance

Each column is one root, for y from 2 to 400; each row is a partial quotient, the first at the top. The brighter a cell, the larger the quotient, and a large quotient means the convergent before it is an unusually good fraction. Near 27 and 256 the bright cells pile up at the top: there, the whole number is the fraction.

4 27 256 1 30 y, from 2 to 400 · rows: the partial quotients, the first at the top a = 2 3–5 6–20 21–100 over 100
Every cell is computed: the quotients of the computed roots, kept where the 120- and 160-digit roots agree; a missing cell is a quotient the two did not agree on. A quotient of 1, the commonest, is the ground. The roots at 4, 27 and 256 have a single quotient, the whole number itself, and leave an empty column.Record: corpus/claimbox-xx/records/xxy-study.json · sha256 bf49436db85dThe build refuses this figure if a whole-number root’s column holds a quotient, or a computed root other than kᵏ is whole.

The roots, in detail

ythe root xclosest fraction, q ≤ 1,000its errorthe bracketboth sides: here · claimbox
21.559610469462…641/4112.361337e-07471/302 < x < 641/411CERTIFIED CERTIFIED
31.825455022925…502/275-4.774703e-07502/275 < x < 387/212CERTIFIED CERTIFIED
422/10x ≥ 2 and x ≤ 2CERTIFIED CERTIFIED
102.506184145589…203/81-1.130608e-05203/81 < x < 198/79CERTIFIED CERTIFIED
2733/10x ≥ 3 and x ≤ 3CERTIFIED CERTIFIED
643.399121540052…775/2281.266965e-06758/223 < x < 775/228CERTIFIED CERTIFIED
1003.597285023540…795/2214.433289e-08259/72 < x < 795/221CERTIFIED CERTIFIED
2553.998359701237…2439/6109.545011e-072435/609 < x < 2439/610CERTIFIED CERTIFIED
25644/10x ≥ 4 and x ≤ 4CERTIFIED CERTIFIED
2574.001633623636…2449/6123.632923e-074 < x < 2449/612CERTIFIED CERTIFIED

The root, the closest fraction and its error are the proposer’s, computed. The bracket is decided here, both sides; the first word is this page’s and the second claimbox’s record, each the same on both sides.

decided
exact arithmetic backs it
computed
float, or a number someone printed; nothing decided it
refused
no mark — the void is drawn
why these numbers stand as they do
  • decided — the computed roots less the decided ones
  • computed — computed by the proposer’s Newton iteration in 160-digit arithmetic; nothing decided it
  • decided — a whole number, decided from both sides here
  • decided — zero: the closest fraction is k, and the root is k, decided from both sides
§3 · what the fractions show

Four things the continued fractions say

Read from the computed roots. The brackets are decided; the patterns below are observations on computed numbers, and each is labelled as one. The first is a theorem, cited.

One: only kᵏ has a fraction for a root (a theorem, cited)

The root is a whole number exactly at y = 4, 27, 256 and 3,125, and each is pinned from both sides: x ≥ k and x ≤ k, both certified here. For every other whole y the root is not even algebraic. If it were a fraction p/q in lowest terms with q above 1, then pp = yq·qp, which q cannot divide; if it were an irrational algebraic number, xˣ would be transcendental by the Gelfond–Schneider theorem. That theorem is cited, not decided here, which is why the claim of §1’s last row is refused. So 4 roots are fractions, and every other one is a number no fraction will ever equal.

Two: near kᵏ, one enormous step (an observation)

Just off kᵏ the root sits just off k, by about 1/(kᵏ(1 + ln k)), the reciprocal of the slope of xˣ at k. Its continued fraction says so with one huge quotient: at y = 255 it is 608, where the slope predicts 610.9; at y = 3,126 it is 8,155, against 8154.5. The closest fraction there is the whole number k.

10 100 1,000 10,000 y = 26 y = 28 y = 255 y = 257 y = 3,124 y = 3,126
Bars: the large partial quotient each root shows, computed. Dashed ticks: kᵏ·(1 + ln k), the slope of xˣ at k, computed in floating point. The x axis is logarithmic.Record: corpus/claimbox-xx/records/xxy-study.json · sha256 bf49436db85dThe build refuses this figure if a row of the figure is not a root beside kᵏ.

Three: elsewhere, the roots look like typical numbers (an observation)

Pool the partial quotients of all the irrational roots and they follow the law that almost every real number obeys: a quotient of 1 occurs 0.41162 of the time, against the law’s 0.41504. Their geometric mean is 2.6729, against Khinchin’s constant 2.6855, and (ln qn)/n averages 1.1777, against Lévy’s 1.1866. These are observations on 1,002 computed roots, not a proof.

0.0 0.1 0.2 0.3 0.4 1 2 3 4 5 6 7 8 9 10 the partial quotient a
Bars: how often each partial quotient occurs among the 38,922 pooled from the irrational roots, computed. Dashed ticks: the Gauss–Kuzmin law, log₂(1 + 1/(a(a + 2))), as the study prints it.Record: corpus/claimbox-xx/records/xxy-study.json · sha256 bf49436db85dThe build refuses this figure if a quotient other than 1 is the most frequent.

Four: a few roots have fractions that are almost too good (an observation)

A large partial quotient makes the convergent just before it unusually close. The largest found is 4,651,670, in the root for y = 709. Its convergent sits closer to the root than a fraction of its size usually can. For y = 100, the fraction 795/221 is already off by only 4.433289e-08.

ypositionpartial quotient (computed)the convergent just before it (computed)
709224,651,67020921543179/4735484816
4193235,36213411503747543/3189935931869
6212624,682433820684589981/99395501124953
1012916,855927575580554664451/257542007895192516
441288,503216551952620850577/51251230167401757
3,12618,1555/1
decided
exact arithmetic backs it
computed
float, or a number someone printed; nothing decided it
refused
no mark — the void is drawn
why these numbers stand as they do
  • computed — a partial quotient of the computed root, kept where the 120- and 160-digit roots agree; computed, not decided
  • computed — kᵏ(1 + ln k) in floating point, as the study prints it
  • computed — a statistic of the computed continued fractions; an observation, nothing decided it
  • computed — the law’s value as the study prints it, rounded to nearest; a literature value, not decided here
  • computed — computed by the proposer’s Newton iteration in 160-digit arithmetic; nothing decided it
  • decided — the position in the continued fraction, counted from the computed quotients
§4 · how it was done

Proposed, stated, decided

What the decisions rest on. Each verdict above is an exact integer comparison, and the step from the root to that comparison is decided: proved in Lean 4 with Mathlib (lean/xx-reduction). For a whole N ≥ 2, xˣ = N has exactly one positive real root (XXReduction.existsUnique_root), and for whole numbers 1 ≤ q ≤ p it lies above p/q exactly when pp < Nq·qp (XXReduction.lt_root_iff; XXReduction.le_root_iff, XXReduction.root_lt_iff and XXReduction.root_le_iff for the other three comparisons, XXReduction.natCast_le_root_iff and XXReduction.root_le_natCast_iff for a whole number). Every one of the 15 theorems passes Lean’s kernel inside lake build, Lean’s kernel again inside Comparator on a fresh runner, and nanoda, a kernel written independently of Lean’s (run 41221355209), and depends on no axiom beyond propext, Classical.choice and Quot.sound. A fraction the study proposed is above 1 in lowest terms, so 1 ≤ q < p; a whole-number bound is q = 1: every comparison on this page is inside the theorems. What is left is read or cited:

  1. assumed
    Only a positive x counts as a root the theorems speak of positive reals (Mathlib’s Real.rpow); what xˣ means for a negative x, and that no negative x is a root worth counting, is the convention claimbox’s library states (claimbox/library/selfpower.py, copied here); read, not decided
  2. assumed
    The recorder computes pp and Nq·qp correctly Python’s whole numbers in tools/run-xx-ledger.py; the theorems say what the comparison means, not that the code computes it. Tested by 228 red controls, not proved
  3. assumed
    Only kᵏ has a fraction for a root (§3, one) the Gelfond–Schneider theorem, cited, with the two-line divisibility argument for a fraction; not decided here, and the claim of §1’s last row is refused for that reason

What is not claimed. That the computed digits are right beyond the decided ones; that the observations of §3 are theorems; that the root of 64 is transcendental (true, cited, refused here); anything about a y the study did not decide. The study’s numbers are a proposer’s, and this page draws them dashed.

decided
exact arithmetic backs it
computed
float, or a number someone printed; nothing decided it
refused
no mark — the void is drawn

Re-run: python3 -I tools/run-xx-ledger.py --check · make drift (the copies against claimbox) · in claimbox: python3 -m claimbox check corpus/xxy/claims.claim --run · the Lean proof: .github/workflows/xx-reduction.yml on GitHub Actions, recorded by node tools/record-xx-reduction.js <run>.

why these numbers stand as they do
  • decided — the theorems of the Lean record, each PROVED
contact

Write to Carlos

Sent by Web3Forms to Carlos’s inbox; nothing stored here.

Or write to carlos@carlostoledo.co, or open the public pilot form on GitHub.